AI Narrative · Web · In Development

Fragile

A Façade-inspired, turn-based interactive mystery steered by a trainable Drama Manager.

Team: Qinchuan Zhang
Type: Research · AI Narrative · USC CSCI 566
Stack: React · FastAPI · LLMs · Python
Status: In development · 2026

Overview

Fragile is a research prototype for an LLM-driven, trainable Drama Manager, inspired by Façade (Mateas & Stern, 2005), the landmark interactive drama in which a player, typing in natural language, becomes entangled in the collapsing marriage of an AI-driven couple. Façade's achievement has almost never been reproduced, and the reason was never vision but cost: two authors spent roughly two person-years hand-writing every beat, line, and reaction, and its “broad and shallow” parser still misread the player often enough to break the fiction. Fragile tests whether LLMs can lift that authorial bottleneck, in a single scene built on the bones of Hitchcock's Rope: the night before Julian and Mia emigrate, you, an old college friend, come to their farewell dinner, served on a shipping crate in an emptied flat. Leo, the friend they invited, never shows up. Each turn you write what you say and do; the system answers with the next moment of the story in the first person, and the couple keeps performing that nothing happened tonight.

Drama Manager 2.0

At the core is a Drama Manager 2.0 that keeps Façade's most durable abstraction, the beat, a minimal unit of dramatic action chosen against an Aristotelian tension arc, while replacing its brittle internals. A frozen LLM proposes beats and, when nothing on the shelf fits, generates a new one; a trainable scorer chooses among the eligible beats, learning the reward function that Façade hand-coded as a formula. Inside a beat, the LLM performs the scene live: the beat fixes the dramatic intent and its effects on the story state, while the dialogue flexes to whatever the player actually says. The four learned modules (beat scoring, input understanding, consistency checking, tension estimation) each run on a rule path, an LLM path, or a trained path, and can shadow one another, so a trained model is compared against the others on the same live turns before it takes over.

One turn

A turn runs through a fixed pipeline. The player's free text is first read by an input reviewer that splits it into speech, action, and expression, maps actions onto a closed table, and refuses what a player cannot decide (another character's actions, the outcome of an attempt, a convenient new object), each time offering a rewrite. Deterministic world rules then resolve where everyone is and who can hear you: the flat is a real floor plan, and a whispered word on the balcony never reaches the kitchen. Characters react within bounded steps of composure and suspicion; the beat runner delivers the beat's next steps; the Drama Manager picks what comes next when a beat ends; and a guard checks every passage the narrator writes against what the player has actually learned, rewriting it rather than let a secret slip. Every decision, with all its candidates and their scores, is logged for training.

Characters

Julian and Mia stay on stage the whole evening, each pulling the player into their own version of the day; Nora, Leo's younger sister, arrives late, first as messages, then a call, then a knock at the door. Every character keeps layered secrets, small ones told to cover larger ones, and the ending follows from what the player accumulated: whose side they took, how much came out, who broke first, and what they told Nora.

Learning from play

The scorer learns from two sources. Human annotators make prefix-blind pairwise judgements: shown the story so far but never what followed, they pick the better of two candidate beats at the same point, so a judgement is not biased by how the session happened to end. At scale, LLM players play whole sessions against the engine: each player gets a personality sampled from the HEXACO model with published human trait correlations, turned into concrete behaviour rules and a goal for the evening, and the player models rotate between providers so the data does not collapse into one model's habits. An LLM judge, calibrated against the human labels, turns these sessions into training pairs for the scorer.